[amy-why-3459]

Purpose

Update the MiniCPM-o 4.5 A3 mean TTFT baseline for concurrency 8 / 128 requests, as requested for #9095.

Metric	Previous baseline	Reported measurement	New baseline
A3 C8 mean_ttft_ms	506.3193 ms	567.9077 ms	624.6985 ms

The new baseline is 567.9077 × 1.10, rounded to four decimal places. The measurement comes from issue #9095, dated 2026-10-18, using vLLM 0.31.0 and Omni 72a665bbe8b4fb035baf56516d643b7fd1f6b039. Add that provenance to the configuration description.

Only this baseline and its description change. This is a baseline adjustment, not a performance optimization or a new benchmark result. Runtime behavior, other metrics/concurrencies, and the comparison policy are unchanged, so serving design contracts and user guides do not change.

Contributor intent

Fix #9095

• Outcome: adjust the A3 C8 TTFT baseline associated with [Bug][Perf]: minicpm, performance metrics regressed by more than 10% compared to the baseline in some scenarios #9095.
• Variable: the mean_ttft_ms baseline for concurrency 8 / 128 requests.
• Change: set it to the reported measurement plus 10%.
• Need: the requested performance-baseline adjustment for the scenario reported in [Bug][Perf]: minicpm, performance metrics regressed by more than 10% compared to the baseline in some scenarios #9095.

Validation

• Parsed the JSON and checked the C8 / N128 index and arithmetic.
• Verified that all other configuration values are unchanged.
• git diff --check: passed.
• pre-commit run --files tests/dfx/perf/tests/test_minicpmo_4_5.json: all applicable checks passed.
• Diff-scoped self-review: one configuration file, no runtime code or new examples.

No new NPU benchmark or full test suite was run; the historical issue measurement is the evidence for the requested baseline update.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-26T04:22:29.470771Z We couldn’t complete this review. Comment @codex review to try again.	be9da57	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[amy-why-3459]

@yenuo26 @natureofnature PTAL

[vllm-omni-review-bot]

This PR appears to be related to model: minicpm.

Model owners: @y-null @yenuo26 @NickCao

Routing: @y-null via semantic router, model owner; @yenuo26 via CI owner, CODEOWNERS; @NickCao via CODEOWNERS

@amy-why-3459, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[natureofnature]

Checked the complete JSON diff, the C8/N128 sweep mapping, and the measurement in #9095. The A3 value is at index 2, 567.9077 × 1.10 rounds to 624.6985, and all other configuration values are unchanged. The current pre-commit and Python build checks pass. This review covers the baseline adjustment; the cited measurement is historical and does not establish current-head serving performance.

【Conclusion: Found no blocking issues in commit be9da57.】
【Action recommended: None.】

[Gaohan123]

What is the reason of relaxing the threshold?

[natureofnature]

@Gaohan123 Compared with the existing baseline, the Oct 18 A3 C8 run shows 35% lower first-audio latency, 49% lower RTF, and nearly doubled request throughput, at the cost of a 12% increase in text TTFT (about 62 ms). For this speech-output workload, that is a reasonable tradeoff and provides a rationale for updating the text-TTFT baseline.

These results reflect the combined pipeline changes from #7929, #8908, and #9026. This PR updates the baseline to reflect the resulting audio-serving performance.

Audio TTFP, RTF, and throughput should be the primary performance criteria for this workload. Text-response latency can be monitored separately with a streaming text-output workload such as Daily-Omni, using its own TTFT baseline and accuracy checks.